Data coherency is threatened whenever two or more computer processes compete for a common data item that is stored in a memory or when two or more copies of the same data item are stored in separate memories and one item is subsequently altered. Previously known apparatus and methods for ensuring data coherency in computer systems are generally referred to as mutual-exclusion mechanisms.
A variety of mutual-exclusion mechanisms have evolved to ensure data coherency. Mutual-exclusion mechanisms prevent more than one computer process from accessing and/or updating (changing) a data item or ensure that any copy of a data item being accessed is valid. Unfortunately, conventional mutual-exclusion mechanisms degrade computer system performance by adding some combination of procedures, checks, locks, program steps, or complexity to the computer system.
Advances in processor and memory technology make it possible to build high-performance computer systems that include multiple processors. Such computer systems increase the opportunity for data coherency problems and, therefore, typically require multiple mutual-exclusion mechanisms.
When multiple processors each execute an independent program to accelerate system performance, the overall system throughput is improved rather than the execution time of a single program.
When the execution time of a single program requires improvement, one way of improving the performance is to divide the program into cooperating processes that are executed in parallel by multiple processors. Such a program is referred to as a multitasking program.
Referring to FIG. 1A, multiple computers 10 are interconnected by an interconnection network 12 to form a computer system 14. FIG. 1B shows than a typical one of computer 10 includes N number of processors 16A, 16B, . . . , and 16N (collectively "processors 16"). In computer 10 and computer system 14, significant time is consumed by intercommunication. Intercommunication is carried out at various levels.
In computer 10, at a processor memory interface level, processors 16 access data in a shared memory 18 by transferring data across a system bus 20. System bus 20 requires a high-communication bandwidth because it shares data transfers for processors 16. Computer 10 is referred to as a multiprocessor computer.
In computer system 14, at an overall system level, computers 10 each have a shared memory 18 and interconnection network 12 is used only for intercomputer communication. Computer system 14 is referred to as a multicomputer system.
The high threat to data coherency in multicomputer and multiprocessor systems is caused by the increased competition among processors 16 for data items in shared memories 18.
Ideally, multicomputer and multiprocessor systems should achieve performance levels that are linearly related to the number of processors 16 in a particular system. For example, 10 processors should execute a program 10 times faster than one processor. In a system operating at this ideal rate, all processors contribute toward the execution of the single program, and no processor executes instructions that would not be executed by a single processor executing the same program. However, several factors including synchronization, program structure, and contention inhibit multicomputer and multiprocessor systems from operating at the ideal rate.
Synchronization: The activities of the independently executing processors must be occasionally coordinated, causing some processors to be idle while others continue execution to catch up. Synchronization that forces sequential consistency on data access and/or updates is one form of mutual exclusion.
Program structure: Not every program is suited for efficient execution on a multicomputer or a multiprocessor system. For example, some programs have insufficient parallelism to keep all multiple processors busy simultaneously, and a sufficiently parallel program often requires more steps than a serially executing program. However, data coherency problems increase with the degree of program parallelism.
Contention: If processor 16A competes with processor 16B for a shared resource, such as sharable data in shared memory 18, contention for the data might cause processor 16A to pause until processor 16B finishes using and possibly updating the sharable data.
Any factor that contributes to reducing the ideal performance of a computing system is referred to as overhead. For example, when processors 16A and 16B simultaneously request data from shared memory 18, the resulting contention requires a time-consuming resolution process. The number of such contentions can be reduced by providing processors 16 with N number of cache memories 22A, 22B, . . . , and 22N (collectively "cache memories 22"). Cache memories 22 store data frequently or recently executed by their associated processors 16. However, processors 16 cannot efficiently access data in cache memories 22 associated with other processors. Therefore, cached data cannot be readily transferred among processors without increased overhead.
Incoherent data can result any time data are shared, transferred among processors 16, or transferred to an external device such as a disk memory 24. Thus, conventional wisdom dictates that computer performance is ultimately limited by the amount of overhead required to maintain data coherency.
Prior workers have described various mutual-exclusion techniques as solutions to the data coherence problem in single and multiprocessor computer systems.
Referring to FIG. 2, Maurice J. Bach, in The Design of the UNIX Operating System, Prentice-Hall, Inc., Englewood Cliffs, N.J., 1986 ("Bach") describes single processor and multiprocessor computer implementations of a UNIX.RTM. operating system in which a process 30 has an asleep state 32, a ready to run state 34, a kernel running state 36, and a user running state 38. Several processes can simultaneously operate on shared operating system data leading to operating system data coherency problems. Bach solves the operating system data coherency problem by allowing process 30 to update data only during a process state transition 40 from kernel running state 36 to asleep state 32. Process 30 is inactive in asleep state 32. Data coherency is also protected by using data "locks" that prevent other processes from reading or writing any part of the locked data until it is "unlocked."
Referring again to FIG. 1, Lucien M. Censier and Paul Feautrier, in "A New Solution to Coherence Problems in Multicache Systems," IEEE Transactions on Computers, vol. C-27, No. 12, December 1978 ("Censier and Feautrier") describe a presence flag mutual-exclusion technique. This technique entails using data state commands and associated state status lines for controlling data transfers among N number of cache memory controllers 50A, 50B, . . . , and 50N (collectively "cache controllers 50") and a shared memory controller 52. The data state commands include shared read, private read, declare private, invalidate data, and share data. Advantageously, no unnecessary data invalidation commands are issued and the most recent (valid) copy of a shared data item can quickly be found. Unfortunately, cache memory data structures must be duplicated in shared memory controller 52 and costly nonstandard memories are required. Moreover, performance is limited because system bus 20 transfers the data state commands.
A. J. van de Goor, in Computer Architecture and Design, Addison-Wesley Publishers Limited, Workingham, England, 1989, classifies and describes single path, single data ("SPDS"), single path, multiple data ("SPMD"), multiple path, single data ("MPSD"), and multiple path, multiple data ("MPMD") computer system architectures from a data coherence perspective as summarized in Table 1.
TABLE 1 ______________________________________ Single Data Multiple Data ______________________________________ Single Path SPSD SPMD Multiple Path MPSD MPMD ______________________________________
Single data ("SD") indicates that only one copy of a data item exists in computer 10, whereas multiple data ("MD") indicates that multiple copies of the data item my coexist, as commonly happens when processors 16 have cache memories 22.
Single path ("SP") indicates that only one communication path exists to a stored data item, whereas multiple paths ("MP") indicates that more than one path to the same data item exists, as in a multiprocessor system with a multiport memory. The classification according to Table 1 results in four classes of computer systems.
SPSD systems include multiprocessor systems that time-share a single bus or use a crossbar switch to implement an interconnection network. In such systems, the processors do not have cache memories. Such systems do not have a data coherence problem because the conventional conditions necessary to ensure data coherence are satisfied--only one path exists to the data item at a time, and only one copy exists of each data item.
Although no processor-associated cache memories exist, shared memory 18 can include a performance-improving shared memory cache that appears to the processors as a single memory. Data incoherence can exist between the shared memory cache and the shared memory. In general, this solution is not attractive because the bandwidth of a shared cache memory is insufficient to support many processors.
MPSD systems are implemented with multiport memories in which each processor has a switchable dedicated path to the memory.
MPSD systems can process data in two ways. In a single-access operation, memory data are accessed sequentially, with only a single path to and a single copy of each data item existing at any time, thereby ensuring data coherence. In a multiple-access operation, a data-accessing processor locks its memory path or issues a data locking signal for the duration of the multiple-access operation, thereby ensuring a single access path for the duration of the multiple-access operation. Data coherence is guaranteed by the use of locks, albeit with associated overhead.
MPMD systems, such as computer 10 and computer system 14, are typically implemented with shared memory 18 and cache memories 22. Such systems have a potentially serious data coherence problem because more than one path to and more than one copy of a data item my exist concurrently.
Solutions to the MPMD data coherence problem are classified as either preventive or corrective. Preventive solutions typically use software to maintain data coherence while corrective solutions typically use hardware for detecting and resolving data coherence problems. Furthermore, corrective solutions are implemented in either a centralized or a distributed way.
Preventive solutions entail using software to designate all shaftable and writable data as noncachable, making it accessible only in shared memory 18. When accessed by one of processors 16, the shared data are protected by software locks and by shared data structures until relinquished by the processor. To alleviate the obvious problem of increased data access time, the shared data structures may be stored in cache memory. The prevention software is responsible for restoring all updated data at shared memory 18, before releasing the software locks. Therefore, processors 16 need commands for purging data from associated cache memories 22.
Unfortunately, preventive solutions require specialized system software, a facility to identify sharable data, and a correspondingly complex compiler. Additionally, system performance is limited because part of the shared data are not cached.
Corrective solutions to data coherence problems are advantageous because they are transparent to the user, albeit at the expense of added hardware.
A typical centralized solution to the data coherence problem is the above-described presence flag technique of Censier and Feautrier.
In distributed solutions, cache memory controllers 50 maintain dana coherence rather than shared memory controller 52. Advantages include reduced bus traffic for maintaining cache data states. This is important because, in a shared bus multiprocessor system, bus capacity often limits system performance. Therefore, approaching ideal system performance requires minimizing processor associated bus traffic.
Skilled workers will recognize that other solutions exist for maintaining data coherence including dedicated broadcast buses, write-once schemes, and data ownership schemes. The overhead cost associated with various data coherence techniques is described by James Archibald and Jean-Loup Baer, in "Cache Coherence Protocols: Evaluation Using a Multiprocessor Simulation Model," ACM Transactions on Computer Systems, vol. 4, No. 4, November 1986. Six methods of ensuring cache coherence are compared in a simulated multiprocessor computer system. The simulation results indicate that choosing a data coherence technique for a computer system is a significant decision because the hardware requirements and performance differences vary widely among the techniques. In particular, significant performance differences exist between techniques that distribute updated data among cache memories and techniques that update data in one cache memory and invalidate copies in other cache memories. Comparative performance graphs show that all the techniques impose an overhead-based performance limit on the computer system.
What is needed, therefore, is a substantially zero-overhead mutual-exclusion mechanism that provides for concurrently reading and/or updating data while maintaining data coherency. Such a mechanism would be especially useful if it is capable of maintaining the coherency of data shared throughout a networked multicomputer system.